The adoption of mmWave radar for human sensing has gained significant attention due to its efficiency, robustness to environmental conditions, and privacy-preserving nature. In particular, gait recognition using radar point clouds presents new opportunities for unobtrusive and resilient biometric authentication and activity analysis. Unlike more traditional representations employed in the radar sensing literature (e.g. micro-Doppler signatures), point cloud data is well-suited for edge computing applications due to its compactness, but pose new challenges due to their noisy and sparse nature. This dataset was collected to build and validate an original neural network model for open-set gait recognition (OSGR) from sparse radar point clouds. The dataset comprises approximately 5 hours of radar measurements collected from 10 subjects with diverse physical characteristics (sex, height, and weight) to ensure a realistic variety of walking patterns. The captures were recorded in a 7.81 × 7.26 m indoor environment, where each participant was instructed to walk freely along random trajectories. Measurements are provided both in point cloud and micro-Doppler spectrogram modalities. To enhance diversity and realism, gait recordings were performed with three different walking manners: Walking freely Walking while holding a smartphone Walking with hands in pockets Each subject was recorded for approximately 30 minutes, with 10 minutes per walking condition, providing a rich and balanced variability in gait patterns. All the measurements were collected with a commercial Texas Instruments MMWCAS-RF-EVM FMCW Radar, operating in the 77-81 GHz frequency band at a frame rate of 10 Hz. This dataset is intended to support research in gait analysis, human activity recognition, and radar-based sensing applications. For additional details and information about the dataset, the user is redirected to our paper "Open-Set Gait Recognition from Sparse mmWave Radar Point Clouds", available at this link. If you find our work and implementation useful for your research, please cite us using the following BibTeX: @ARTICLE{11080220, author={Mazzieri, Riccardo and Pegoraro, Jacopo and Rossi, Michele}, journal={IEEE Sensors Journal}, title={Open-Set Gait Recognition from Sparse mmWave Radar Point Clouds}, year={2025}, volume={}, number={}, pages={1-1}, keywords={Point cloud compression;Radar;Feature extraction;Millimeter wave communication;Training;Gait recognition;Neural networks;Legged locomotion;Inference algorithms;Deep learning;Point cloud;mmWave Radar;Open Set Classification;Deep Learning}, doi={10.1109/JSEN.2025.3587503}}

The dataset includes radar measurements both in point-cloud and micro-doppler spectrogram formats. Point Clouds:Point cloud files are organized with the following folder structure:dataset_dir/point_clouds///pc__.obj Each .obj file is a serialized Python object, loadable with the pickle module, containing a point cloud sequence capture of approximately 1 minute in the form of a list of dictionaries, each of which containing details about each point cloud. Micro-doppler spectrograms:Micro-doppler spectrogram files are organized with the following folder structure:dataset_dir/spectrograms///md__.ptIn this case, each .pt file is a PyTorch tensor, easily loadable using the torch.load() method, also comprising approximately 1 minute of measurements. Differently from point clouds, where each point cloud could have a variable number of points, spectrograms are easily stored in tensor format. For detailed code examples on how to handle the dataset, the user is redirected to the official code repository of our paper, available at this link.